This repo is implemented based on detectron2 and CenterNet

• Support imgaug data augmentation
• Support swa
• Support Knowledge Distill, teacher-student, designed by myself
• Support other LR_SCHEDULER
• Support Optimizer RangerLars, not convergence in COCO
• We provide some examples and scripts to convert centerX to Caffe, ONNX and TensorRT format in projects/speedup

• [️✔] Support simple inference
• [✔] Support to caffe, onnx, tensorRT
• Support keypoints

• Python >= 3.7
• PyTorch >= 1.5
• torchvision that matches the PyTorch installation.
• OpenCV
• pycocotools

pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

• GCC >= 4.9

gcc --version

• detectron2

pip install -U 'git+https://github.com/CPFLAME/detectron2.git'

• pytorch tools

pip install git+https://github.com/pabloppp/pytorch-tools -U

the same as detectron2

modify your yamls in run.sh

sh run.sh

modify your yamls in run.sh, add eval-only and MODEL.WEIGHTS in your setting

sh run.sh

This repo use less training time to get a competitive performance compared to other versions

Backbone ResNet-50

Backbone ResNet-18

centerX_KD means ResNet-50(33.2) as teacher, ResNet-18(27.9) as student, Knowledge Distill for 70 epoch in coco.

• optim: SGD
• lr: 0.02
• scheduler: WarmupMultiStepLR ,drop 0.1 in (50, 62) for 80 epoch; (90 ,120) for 140 epoch
• train size: 512 max size
• test size: 512 max size
• batch size: 64
• woGT: KD only use teacher loss

Generalization performance for Knowledge Distill

centerX is released under the Apache 2.0 license.

• detectron2
• CenterNet
• CenterNet-better
• CenterNet-better-plus
• FastReID